Building realistic voice interactions with voice AI

Content

Messy. A lot of parameters to tune to cater to a variety of end user conversation styles.

Here is some of the settings I added to Talentprism to lend more control over AI behavior to the users.

Personality of the interviewer

LLM model - this is the biggest factor the determines personality. llama models are more verbose and friendly, the open AI and google models is more succinct but speak more when needed. I would recommend trying the google models (flash 2.0)
Temperature - recommend around 0.7-1 for sufficient randomness in AI responses.
AI Voice - recommend trying Paige or Ana for a very realistic natural voice

Reduce interruptions

Start Speaking Delay -Default is 0.4 secs. If the user is a slow speaker, increase this to 0.6 seconds. This induces a delay in every AI response (like 2 slow speakers talking to each other).

Handle Interruptions

Interruption Sensitivity (default 0.2 sec) and Speech Backoff Time (default 1 sec) - If candidate and AI speak at the same time, the AI stops after 0.2 sec and waits for 1 sec before responding.

Extended silence

Idle Timeout (default 10 sec) - if the candidate is silent for certain duration, the AI will say “are you still there” & “prolonged silence will”
Silence Timeout is when the call will disconnect. default is 30 sec.

Other configurations

You can edit other config like the first and last AI response when the call starts and ends, max call duration, audio or video call, welcome message, etc.